acceptodds
Under review as a conference paper at ICLR 2027

Neuron-Aligned Model Merging for Cross-Modal Reasoning Transfer

Abstract

Existing neuron-level model merging methods transfer reasoning capabilities from LLMs to VLMs by selectively merging reasoning-relevant neurons at the same neuron indices. However, reasoning training and multimodal adaptation can reshape neuron responses, breaking the same-index correspondence assumption: same-index neurons may become functionally mismatched, while cross-index correspondences may be overlooked. To address this limitation, we propose Targeted Alignment and Neuron Grafting via Optimal Transport (TANGO), a training-free framework for cross-modal reasoning transfer through neuron-aligned model merging. TANGO uses activations on shared inputs and optimal transport to establish functional correspondences between source and target neuron spaces, rather than relying on fixed index-wise matches. The resulting correspondences guide both the selection of reasoning-relevant transfer locations and the alignment of source parameters for selective merging. Across four VLMs from the Qwen and InternVL families, TANGO consistently improves multimodal mathematical reasoning and general multimodal understanding, achieving the best or tied-best result in 27 of 28 mathematical reasoning settings and ranking first in 18 of 20 general multimodal and text-only reasoning comparisons.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
model merging, vision-language models, optimal transport, reasoning transfer

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_VJ7NOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TANGO, a training-free framework for transferring reasoning capabilities from LLMs to VLMs through neuron-aligned model merging. Instead of relying on same-index neuron correspondence, TANGO uses shared-input activations and optimal transport to identify functional correspondences between source and target neurons. The learned transport plan guides both task-relevant neuron selection and source-parameter alignment for selective merging. Experiments across Qwen and InternVL models show consistent improvements on multimodal mathematical reasoning and general reasoning benchmarks.

Critical strengths
  1. The paper identifies a reasonable limitation of index-wise neuron merging, as neurons with the same indices may no longer have matching functions after different training and adaptation processes.
  2. TANGO combines activation-based neuron alignment, task-aware neuron selection, and selective parameter merging into a unified framework.
  3. The method achieves the best average mathematical reasoning performance across four VLMs and shows generally positive results on broader multimodal and text-only benchmarks.
  4. The paper evaluates alignment strategies, soft versus hard transport, and several hyperparameters. The controlled permutation experiment also provides evidence that OT can recover known neuron correspondences.
Critical weaknesses
  1. TANGO largely combines existing activation-based optimal transport alignment with task-aware neuron selection and selective model merging. While the integration is reasonable, the paper does not sufficiently establish what fundamentally distinguishes the proposed framework from prior OT-based model merging and neuron-level reasoning transfer methods.
  2. Table 3 shows that perturbing the selected parameters severely degrades model performance, but lacks matched-budget random or low-importance neuron controls. Therefore, the results do not demonstrate that the identified neurons are specifically important for reasoning rather than essential to general model functionality.
  3. The controlled permutation experiment verifies OT alignment under artificial neuron permutations, but does not establish that the learned correspondences are functionally meaningful between independently trained LLMs and VLMs. More direct evidence on real model pairs is needed to support this central assumption.
  4. Table 4 does not fully disentangle OT-based neuron selection from OT-based parameter alignment. Without evaluating these two components independently, it is unclear which contributes most to the reported improvements.
Questions for authors
  1. Beyond the controlled permutation experiment, the paper provides insufficient evidence that the learned OT correspondences are functionally meaningful between independently trained LLMs and VLMs.
  2. The individual contributions of OT-based neuron selection and OT-based weight alignment are not sufficiently isolated, making it difficult to determine their respective effects.
  3. The paper does not adequately investigate how different calibration data distributions affect neuron correspondence and downstream reasoning performance.
  4. The paper lacks a detailed analysis of TANGO's computational and memory overhead.
6 comments

We thank the reviewer for the careful assessment and constructive suggestions.

W1: Distinction from prior OT merging and neuron-level transfer

TANGO combines target-side multimodal relevance uiu_i with OT-transported source reasoning relevance to select neurons. Cui et al. (2026a) selects neurons by activation magnitude in text-only merging. TANGO also replaces SNRF’s same-index matching with OT-based functional alignment. When transferring from DeepSeek-R1-0528-Qwen3-8B to Qwen3-VL-8B-Instruct, removing only uiu_i with the OT plan, merge coefficients, and update budget fixed lowers all five scores.

Selection settingMathVista-CoTMathVerse-TDMMEMMMU-ProGPQA
With visual profiling71.50%57.61%1744.4335.90%40.63%
Without visual profiling69.90%56.98%1722.2834.22%39.51%

W2: Functional importance of selected neurons

Zeroing PG-OT positions lowers all three scores more than matched random zeroing. Both runs start from the same Qwen3-VL-8B-Instruct checkpoint and zero the same number of positions per layer and projection. The zeroing runs use the MME perception score. Original TANGO is shown for reference.

SettingMathVista-CoTMMEGPQA
Original TANGO (no intervention)72.30%1744.4340.63%
Zero TANGO PG-OT positions25.90%4.1723.21%
Zero matched random positions56.40%1596.6732.81%

W3 / Q1: Functional correspondence in real LLM–VLM pairs

Text-derived OT predicts neuron responses when the VLM processes images. In Figure 1, a cross-index pair has Pearson correlation 0.851 versus 0.640 for the same-index pair. Across 36 Q-projection layers, 54.26% of OT mass is off-diagonal. On 500 Geometry3K pairs, the VLM receives images and questions and the LLM receives faithful descriptions and the same questions. Frozen GSM8K OT yields higher weighted correlations than same-index and shuffled matching. Image-based recalibration gives no consistent downstream advantage, supporting shared text as the default.

Calibration schemeMathVista-CoTMathVerse-TD
Original GSM8K text OT72.30%58.76%
Image / faithful-description OT71.10%54.31%
Image / oracle-style-fact OT72.80%55.71%

W4 / Q2: Neuron selection and parameter alignment

Both components contribute. Selection chooses target positions (Eqs. 6–7), and alignment maps source weights to them (Eqs. 8–9). We will clarify these roles in Table 4. The full method scores 67.92 versus 66.38 without OT alignment, 66.91 with shuffled correspondence, and 66.28 with hard transport. Fixing the mask isolates the benefit of OT parameter mapping below. W1 fixes OT and isolates the contribution of uiu_i to selection.

Parameter mappingMathVista-CoTMathVerse-TD
OT parameter alignment72.30%58.76%
Same-index merging71.90%58.25%
Gain from OT alignment+0.40 pp+0.51 pp

Q3: Sensitivity to calibration data

All three corpora yield high held-out correlations, and GSM8K leads on the three downstream benchmarks. We recompute OT, selection, and merging using 500 questions per corpus while fixing the model pair, task profiling, update budgets, and merging coefficients. We test each frozen plan on the same target neurons and held-out sets (100 questions each).

OT calibration corpusMATH held-out PearsonARC held-out Pearson
GSM8K0.84110.857
MATH Level 1–20.85550.8444
ARC-Challenge0.85630.8536
Calibration corpusMMStarMMMU-ProMME
GSM8K62.61%35.90%1744.43
MATH Level 1–262.13%35.55%1738.62
ARC-Challenge62.07%34.20%1737.11

Q4: Compute and GPU memory

On one A100 80GB in BF16, TANGO cuts three-stage compute time versus SNRF by 34.9% on Qwen3-8B and 15.0% on InternVL3.5-8B, with comparable peak memory. Faster merging offsets alignment cost. TANGO and SNRF average 55.94% and 55.06% on Qwen3-8B, and 56.67% and 56.22% on InternVL3.5-8B, respectively, across the mathematical reasoning tasks. The merge runs once offline. Inference uses the original target architecture without the source expert or OT.

BackboneMethodProfiling (s)Alignment / selection (s)Merging (s)Total (s)Peak memory (GiB)
Qwen3-8BTANGO130.7466.346.85203.9232.04
Qwen3-8BSNRF105.750.18207.3313.2332.78
InternVL3.5-8BTANGO546.2569.426.41622.0831.58
InternVL3.5-8BSNRF524.530.21207.23731.9632.33

Dear Reviewer VJ7N,

Thank you for suggesting additional controls for TANGO’s selection and alignment mechanisms. The new experiments show that both design choices improve transfer, and TANGO achieves higher average mathematical reasoning accuracy than SNRF with less offline compute on both tested backbones.

The choice of neurons matters. Zeroing TANGO-selected positions caused greater degradation across all three benchmarks than zeroing the same number of random positions. With OT fixed, multimodal profiling improved all five evaluated metrics. With the selection mask fixed, OT parameter alignment improved both reasoning benchmarks over same-index merging.

Correspondence holds on real model pairs. The frozen text-derived OT plan remained predictive on paired image and description inputs, and all three tested text calibration corpora yielded high held-out response correlations.

Less offline compute. Compared with SNRF, TANGO reduced the combined time for profiling, alignment or selection, and merging by 34.9% on Qwen3-8B and 15.0% on InternVL3.5-8B, with comparable peak GPU memory.

If you have a chance to revisit the rebuttal, we would appreciate your taking these results into account in your final assessment. We would be glad to discuss any concerns that remain.

Best regards, The authors

Thank you to the authors for providing additional experimental results, which address part of my concerns. Regarding W4, I would like to further understand the effectiveness of OT-based parameter alignment in the absence of neuron selection. Specifically, what would the performance be if the selection mask (M) were removed and the OT-aligned source weights were directly merged with the target weights?

Reviewer_7mcEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a cross-modal model merging method based on optimal transport named TANGO. Based on closely related work on a cross-architecture LLM merging method with OT (Cui et al., 2026a), this paper introduces some new techniques to adapt the method for LLM -> VLM tasks. The method is evaluated on widely used backbones and multimodal benchmarks.

Critical strengths
  • The paper overall is easy to follow. The proposed method is well-motivated and a comprehensive experimental study is provided.

  • Cross-modal knowledge transfer is an important and valuable problem.

Critical weaknesses
  • My main concern is the improvements and contributions over the closely related work Cui et al., 2026a. The core idea of using OT for matching neurons in different model architectures is already presented in Cui et al., 2026a. Although this paper proposes a task-aware OT selection, it is still unknown if the change can largely improve the existing solution. The baseline method Cui et al., 2026a is also not compared in experiments. It is necessary to provide direct evidence to show the effectiveness of the proposed changes. The overall technical contributions over the existing solution also look minor, especially considering closely related work like Cui et al., 2026b.

  • In the main experiments, LLM and VLM models share the same basic LLM architecture. How about the results if LLMs and VLMs have different architectures? Will the proposed method support this case?

  • There are some misalignments between Table 3 and Table1/2. In Table 3, Qwen3-VL-8B None (TANGO) achieves 1727.06 on MME, while the same number is the Qwen3-VL-8B Base result in Table 2.

Questions for authors

Please refer to my comments above. Response on W1 could affect my assessment.

After rebuttal: The authors provide important ablation results of the proposed method in the rebuttal. For the cross-architecture transfer, although Qwen3 and InternVL-3.5 belong to different model families, InternVL-3.5 is actually initialized from the Qwen3-8B LLM backbone, which may raise concerns about the generalizability and applicability of the proposed method. Considering the overall technical contribution, I will maintain my initial rating.

4 comments

We thank the reviewer for the careful assessment and constructive feedback. We address the distinction from prior OT merging, the scope of cross-architecture transfer, and the correction to Table 3 below.

W1 & Q: Novelty Beyond Cui et al. (2026a)

TANGO extends OT-guided neuron selection by using multimodal task relevance to decide where to transfer reasoning knowledge. We profile the target VLM on multimodal tasks to obtain a relevance indicator uiu_i and combine it with OT-transported source reasoning relevance and correspondence concentration. This adds a target-side multimodal selection criterion to the text-only LLM merging setting studied by Cui et al. (2026a).

To isolate the effect of this criterion, we transfer from DeepSeek-R1-0528-Qwen3-8B to Qwen3-VL-8B-Instruct and compare selection with and without uiu_i, keeping the OT plan, source relevance, merging coefficients, and per-layer and per-projection update budgets fixed.

Selection settingMathVista-CoTMathVerse-TDMMEMMMU-ProGPQA
With visual profiling (uiu_i)71.50%57.61%1744.4335.90%40.63%
Without uiu_i69.90%56.98%1722.2834.22%39.51%

Retaining uiu_i improves all five metrics. With alignment and update counts fixed, these gains show that multimodal profiling helps select more effective locations for parameter transfer.

W2: Transfer Across Different Architectures

TANGO supports transfer across VLM families when the source LLM and the target VLM’s language tower have compatible layers and projection dimensions. Our Qwen and InternVL experiments illustrate this setting. TANGO aligns and updates parameters in the language tower while preserving the target’s visual architecture. Applying the current method to language backbones with incompatible depths, widths, or projection layouts would require an explicit structural mapping before neuron alignment and parameter transfer.

W3: Correction of Table 3’s TANGO Reference Scores

We have corrected Table 3 in the revised manuscript PDF. The “None (TANGO)” rows now report the unperturbed TANGO scores and match the corresponding results in Tables 1 and 2. In particular, the Qwen3-VL-8B MME score has been corrected from 1727.06 to 1744.43.

ModelMathVista-CoTMathVerse-TDMMEMMMU-ProGPQA
Qwen3-VL-8B72.358.761744.4335.940.63
Qwen2.5-VL-7B71.958.51722.2133.8234.38
InternVL2.5-4B61.343.151685.523.4130.58
InternVL3.5-8B71.147.721705.6630.1742.63

Dear Reviewer 7mcE,

Thank you for asking us to clarify what TANGO adds to prior OT merging. TANGO uses evidence from multimodal tasks to identify which target neurons should receive reasoning parameters. Our new ablation shows that this choice improves transfer even when both variants use the same OT plan and update budget.

Better transfer at the same budget. Retaining target-side multimodal relevance improved all five evaluated metrics, including MathVista-CoT from 69.90% to 71.50% and MathVerse-TD from 56.98% to 57.61%.

Reuse across VLM families. TANGO transfers reasoning parameters into Qwen and InternVL targets while preserving their visual architectures. The current method operates on compatible language towers, and we clarify where additional structural mapping would be needed.

Corrected reference scores. The revised PDF now reports the unperturbed TANGO scores in Table 3. The Qwen3-VL-8B MME entry has been corrected from 1727.06 to 1744.43.

We hope the ablation makes the contribution beyond Cui et al. (2026a) clearer. We would be grateful if you could consider it in your final assessment, and we welcome any remaining questions.

Best regards, The authors

Dear Reviewer 7mcE,

Thank you again for your careful reading and for pressing us to explain what TANGO adds to prior OT-based merging. Your question led us to a more focused experiment. With the OT plan and update budget fixed, including target-side multimodal relevance improved all five evaluated metrics. We are grateful that your feedback pushed us to make this part of the paper more precise.

If you have a moment before the discussion closes, we would be very grateful to hear whether the ablation helps address your main concern. If there is a specific point that still seems unclear or unconvincing, we would welcome the chance to clarify it. Thank you again for the time and care you have given our work.

Best regards,
The authors

Thank you for the additional ablation results. Although Qwen3 and InternVL-3.5 belong to different model families, InternVL-3.5 is actually initialized from the Qwen3-8B LLM backbone, which may raise concerns about the generalizability and applicability of the proposed method. Considering the overall technical contribution, I will maintain my initial rating.

Reviewer_GCKJOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

Neuron-Aligned Model Merging for Cross-Modal Reasoning Transfer investigates how to improve an existing VLM by selectively merging parameters from a text-only reasoning expert without additional training. Existing neuron-level merging methods commonly assume that neurons at the same indices correspond functionally across the source and target models. The paper argues that reasoning training and multimodal adaptation can invalidate this assumption, creating mismatched pairs and overlooking useful cross-index correspondences.

The proposed framework, TANGO, estimates neuron correspondences through entropy-regularized optimal transport over activation responses to shared text-only calibration inputs. Separately, task-specific text and multimodal corpora identify reasoning-relevant source neurons and task-relevant target neurons. The same transport plan maps source relevance into the target space to guide transfer-location selection and aligns source weights before selective interpolation. The method updates selected language-model parameters while keeping the target's vision encoder and multimodal projector fixed.

Across four VLMs from the Qwen and InternVL families, TANGO achieves the best or tied-best result in 27 of 28 multimodal mathematical reasoning settings and ranks first in 18 of 20 general multimodal and text-only reasoning comparisons against the evaluated training-free baselines. Controlled permutation experiments support correspondence recovery, and ablations support the use of learned alignment and soft transport. These findings establish consistent gains within the tested model-reuse setting, while leaving open whether text-derived correspondences generalize to image-conditioned behavior and how much of the improvement comes from downstream reasoning versus better use of visual evidence.

Critical strengths
  • The transfer methodology has a principled mathematical formulation that addresses the limitations of index-based neuron matching. TANGO defines neuron similarity through activation responses to shared inputs and formulates correspondence as entropy-regularized optimal transport with a Pearson-correlation cost (Sections 3.1-3.3). This provides an explicit objective for matching neurons by observed behavior, while allowing soft, distributed correspondences across indices. Reusing the same transport plan to map source relevance scores and align source parameters makes transfer-location selection and parameter transfer internally consistent (Sections 3.3-3.4). The rationale is reinforced by the function-preserving permutation construction in Appendix B.1: corresponding permutations of the FFN projections leave its output unchanged, demonstrating why neuron indices alone cannot identify function. In this controlled setting, TANGO maintains held-out functional correlation close to 1.0 even under complete permutation. Ablations also favor the learned correspondence and soft transport over shuffled alignment and hard matching (Table 4). These results support the formulation and alignment mechanism, although they do not constitute a general theoretical guarantee of successful capability transfer between independently trained models.

  • The method delivers consistent performance gains across multiple VLMs and evaluation domains without additional model training. Across four target models from the Qwen and InternVL families and five training-free transfer baselines, TANGO achieves the best or tied-best result in 27 of 28 multimodal mathematical reasoning settings and the highest mathematical reasoning average for every target model (Table 1). Its average gains over the strongest baseline are 0.90 percentage points for Qwen2.5-VL-7B and 0.88 for Qwen3-VL-8B. The benefits extend beyond mathematical reasoning: TANGO ranks first in 18 of 20 general multimodal and text-only reasoning comparisons and second in the remaining two (Table 2). While the margins are generally modest, their consistency across model families and tasks supports the practical effectiveness of selective, functionally aligned transfer within the evaluated setting.

Critical weaknesses

The practical motivation is limited to a specific model-reuse setting, and the advantage over directly using an existing multimodal reasoning model remains unestablished. TANGO addresses a useful but relatively narrow scenario: a user wishes to retain a particular VLM and has access to a compatible text-only reasoning expert whose capabilities may improve that VLM without additional training. Its improvements over the original VLMs and other training-free merging methods establish effectiveness within this setting (Sections 4.1-4.2). They do not establish when merging is preferable to directly deploying an available multimodal reasoning model. In particular, the Qwen3-VL-8B and InternVL3.5-8B targets use Instruct checkpoints (Appendix Table 5), and the paper does not report comparisons with their Thinking counterparts. Comparisons with similarly sized multimodal reasoning models, accounting for inference budgets and the cost of profiling, alignment, and merging, would clarify the practical benefit. Such comparisons would contextualize the method's utility without negating its contribution as a model-reuse technique.

Questions for authors
  • Do neuron correspondences estimated from text-only inputs remain valid when the VLM processes images, or do they primarily capture textual behavior? Section 3.2 estimates the OT correspondence using shared text-only calibration inputs. Although target-neuron relevance is measured on multimodal tasks, this does not directly establish that the selected source-target pairs retain corresponding functions under image-conditioned inputs. Could the authors test this generalization using matched image-question and faithful textual-description inputs, and measure whether the text-derived correspondence remains predictive of neuron responses and useful for transfer? This would clarify the scope of the claimed functional correspondence beyond the calibration modality.

  • What systematic evidence distinguishes better downstream reasoning from improved use of visual evidence as the source of the gains? Appendix B.4 selects cases where TANGO succeeds and both the original model and SNRF fail. The table-reasoning example in Figure 8 is consistent with improved downstream constraint handling: the original model reads the relevant numerical entries but incorrectly includes values above the specified upper bound, whereas TANGO applies the range correctly. However, selected successes without systematic error statistics cannot establish that this mechanism accounts for most of the improvement, nor do they validate text-derived neuron correspondences under image-conditioned inputs. Could the authors quantify these error types on a representative evaluation set and use controlled comparisons with faithful textual descriptions or oracle-extracted visual facts to separate visual extraction from subsequent reasoning? Improved textual reasoning that benefits multimodal tasks would still be valuable; the requested evidence would clarify the mechanism and the interpretation of cross-modal transfer.

3 comments

We thank the reviewer for the thoughtful feedback and for suggesting comparisons that clarify TANGO’s practical value and how it improves multimodal reasoning.

W1: Practical Value and Efficiency

TANGO is preferable when users want to retain an existing VLM, have a compatible reasoning expert, and need low response latency. It improves the target VLM through a one-time, training-free merge. The merged model runs without the source expert or OT at inference.

Compared with base VLMs, TANGO improves average mathematical reasoning accuracy by 1.30 and 0.82 percentage points (pp) on Qwen3-VL-8B and InternVL3.5-8B, respectively. It achieves the highest average accuracy among the evaluated training-free methods on all four target VLMs.

Compared with Thinking models, TANGO delivers 8.47×8.47\times and 36.91×36.91\times speedups in paired MathVista-CoT timings for Qwen3 and InternVL3.5, respectively. Its benchmark accuracy is lower by 8.80 and 6.40 pp, respectively. This trade-off favors TANGO when response latency is a deployment priority.

BackboneTANGO (s/question)Thinking (s/question)Speedup
Qwen3-VL-8B11.6798.888.47×8.47\times
InternVL3.5-8B3.46127.7536.91×36.91\times

Compared with SNRF, a related sparse merging method, TANGO reduces the combined time for profiling, alignment or selection, and merging by 34.9% and 15.0%, respectively, with comparable peak GPU memory. Faster parameter merging offsets the added alignment cost. These three stages run once before deployment.

BackboneMethodProfiling (s)Alignment / selection (s)Merging (s)Total compute (s)Peak memory (GiB)
Qwen3-8BTANGO130.7466.346.85203.9232.04
Qwen3-8BSNRF105.750.18207.30313.2332.78
InternVL3.5-8BTANGO546.2569.426.41622.0831.58
InternVL3.5-8BSNRF524.530.21207.23731.9632.33

Q1: Cross-Modal Validity of Text-Derived OT

Text-derived OT remains predictive of neuron responses when the VLM processes images and supports effective parameter transfer in the evaluated model pair.

Response prediction. We freeze the original text-calibrated OT plan and give the VLM the image and question while giving the source LLM a faithful description and the same question. The frozen plan yields higher response correlations than shuffled correspondence.

Downstream transfer. We use GPT-6.1-Sol to generate faithful descriptions from images alone and oracle-style visual facts from images and questions. A separate model pass reviews each description. Both steps exclude reference answers and solution steps. We then compare parameter alignment strategies with the target mask, merging coefficients, and update budget fixed.

Calibration / alignmentMathVista-CoTGain over BaseMathVerse-TDGain over Base
Base VLM71.00%—57.99%—
Original text OT72.30%+1.30 pp58.76%+0.77 pp
Image / faithful-description OT71.10%+0.10 pp54.31%−3.68 pp
Image / oracle-style-fact OT72.80%+1.80 pp55.71%−2.28 pp
Same-index alignment71.90%+0.90 pp58.25%+0.26 pp

Text OT improves both benchmarks and outperforms same-index alignment on both. Oracle-style OT gives the largest MathVista gain but falls below Base on MathVerse. These results support shared text as a practical default for OT calibration. It avoids description generation and review, while multimodal profiling guides the choice of target neurons.

Q2: Visual Extraction vs. Downstream Reasoning

TANGO’s advantage persists when both models receive identical visual facts, showing that downstream reasoning or use of the supplied evidence contributes to the gains.

Controlled evaluation. On all 1,000 MathVista-Solution questions, we replace the image with the same model-extracted and separately reviewed oracle-style visual facts for both models. TANGO’s advantage over Base is 3.40 pp with shared facts and 2.30 pp with the original images.

Input conditionBaseTANGOGain
Original image + question71.30%73.60%+2.30 pp
Oracle-style facts + question74.40%77.80%+3.40 pp

Error analysis. Model-assisted labeling of the full response set, with model identities hidden, finds fewer visual perception, downstream reasoning, and mixed errors for TANGO. Together with the shared-fact control, this extends Figure 8 beyond selected cases and identifies downstream reasoning as one contributor to the observed gains.

Dear Reviewer GCKJ,

Thank you for asking us to clarify when TANGO is useful. Our new results show that TANGO improves an existing VLM without additional training while keeping response latency substantially lower than that of same-scale Thinking models.

Accuracy and latency. The Thinking models achieved higher accuracy on the full MathVista-CoT benchmark, but took 8.47× and 36.91× as long to respond on the paired timing subset for Qwen3-VL-8B and InternVL3.5-8B, respectively. TANGO also improves average mathematical reasoning accuracy over both base VLMs, making it a practical option when response time matters.

Text calibration under image inputs. The frozen text-derived OT plan remained predictive when the VLM processed images. Recalibrating with images and descriptions yielded no consistent downstream advantage across the two benchmarks.

Improved reasoning with the same evidence. TANGO’s advantage over Base increased from 2.30 to 3.40 percentage points when both models received identical visual facts on MathVista-Solution, supporting a contribution from downstream reasoning.

If you have a chance to revisit our response, we would appreciate your considering these results in your final assessment. We would be happy to clarify any remaining questions.

Best regards, The authors

Dear Reviewer GCKJ,

Thank you again for such a perceptive review. Your question about when a practitioner would actually choose TANGO over a same-scale Thinking model pushed us to make the deployment trade-off much clearer. Your suggestion to give both models the same visual facts also led to one of the most informative additions to the revision. We are genuinely grateful for both suggestions.

If you have a moment before the discussion closes, we would be especially grateful to hear whether the new comparisons address your main concern about practical value. Even a brief indication of what remains unclear would help us focus any further clarification. Thank you very much for your time and consideration.

Best regards, The authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.